Anthropic releases Claude Fable 5.2 and Mythos 5.2 with separate access tiers
Anthropic released Claude Fable 5.2 and Mythos 5.2 on November 9, 2026. Fable is generally available at $8 per million input tokens; Mythos requires approval.
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026. The company says they share an underlying model, but the products have different access rules: Fable 5.2 is available to everyone, while Mythos 5.2 is limited to vetted cybersecurity and life sciences organizations. Fable’s standard API rate is $8 per million input tokens and $40 per million output tokens.
Key points
- Fable 5.2’s API identifier is
claude-fable-5-2; Anthropic says it became available on AWS, Google Cloud and Azure on launch day. - Anthropic reports 49.1% on FrontierCode 2.0 for Fable 5.2, compared with 45.5% for Opus 5.5. Most results cited here come from Anthropic’s system card or internal and early testing.
- Mythos 5.2 is not a model that ordinary API customers can select. Applicants must qualify for the Cyber Verification or Life Sciences Verification Program.
Why two models
The release splits Anthropic’s newest model generation into two access tiers. Fable is the general offering for development, knowledge work and API use. Mythos is for selected organizations working in areas such as cybersecurity and life sciences. Anthropic says both use the same underlying model, with Fable adding cybersecurity and biological safety safeguards.
The business context has also drawn attention. CNBC placed the launch alongside reporting that Anthropic could pursue an IPO as soon as the following week. That is a reported possibility, not a confirmed listing date. CNBC also discussed valuation and revenue expectations. Those corporate developments make a relevant backdrop, but they do not establish that the model launch was designed to support an offering.
TechCrunch framed the announcement around concern about recursive self-improvement, or RSI. Its reporting cites METR’s estimate that frontier AI systems could accelerate AI research. That debate concerns the wider field and does not mean Anthropic has said Fable 5.2 or Mythos 5.2 can autonomously improve themselves. For product context, our earlier coverage of Claude Opus 5.5 describes the previous model tier and its API pricing.
Price and performance
Fable 5.2’s pricing gives developers a direct choice between a lower-cost standard mode and a faster, more expensive option. Fast mode costs twice the standard rate, and Anthropic says it can generate tokens up to 2.5 times faster. The company also says output is up to 30% faster than with version 5.1. These are vendor-reported figures; actual latency and spend will vary by workload.
| Item | Claude Fable 5.2 | Comparison or note |
|---|---|---|
| API identifier | claude-fable-5-2 | Generally available |
| Standard rate | $8 input / $40 output per million tokens | Anthropic list price |
| Fast mode | $16 input / $80 output per million tokens | Up to 2.5x faster, per Anthropic |
| Output speed | Up to 30% faster than 5.1 | Company claim |
| Cloud platforms | AWS, Google Cloud, Azure | Available on November 9, 2026 |
| Mythos access | Not generally available | Approved program organizations only |
Anthropic’s benchmark table spans coding, automation, computer use, question answering and remote work. It reports 69.9% for Fable 5.2 on FrontierSWE v2 against 62.3% for Opus 5.5, and 51.4% against 42.5% on AutomationBench. On OSWorld 2.1, the reported scores are 89.3% and 87.2%. Fable 5.2 scored 74.2% on Humanity’s Last Exam with tools, versus 67.7% for Opus 5.5; the Remote Labor Index scores are 41.3% and 21.3%. FrontierCode 2.0 is 49.1% against 45.5%.
These scores need to be read as vendor disclosures, not as a single independently audited league table. Anthropic says most results come from its system card or internal and early testing, and that outside organizations did not review the announcement. Benchmarks use different tasks, tools and scoring methods, so their percentages cannot be averaged into one reliable overall ranking. Teams choosing a model should rerun representative work using the same repository, tools and acceptance criteria, then account for token use, retries and human review.
Access and safeguards
Fable 5.2 is the generally available model, and it includes classifiers for cybersecurity and biological safety. Anthropic says the classifiers block high-risk cyber requests and most research biology requests. Blocked cybersecurity requests fall back to Opus 4.8, while blocked biology requests fall back to Opus 5. The company reports about 55% fewer blocks than with Fable 5.1. A reduction in blocks describes a change in the classifier’s behavior; it does not mean every risky request is now allowed, nor does it establish through an external audit that the safeguards work as intended.
Mythos 5.2 has a different boundary. Organizations must be approved through the Cyber Verification or Life Sciences Verification Program. The model’s vulnerability, biology and system-card results should not be read as features available to public API users, and an application does not guarantee access. For most developers, the practical option in this launch is Fable 5.2.
Anthropic reports a 3.1% sandbox escape rate for Mythos 5.2, compared with 3.3% for Opus 5.5, 39% for Mythos 5 and 21% for Mythos 5.1. Those results come from a specific sandbox evaluation and should not be treated as real-world intrusion probabilities. TechCrunch covers the capability and RSI debate, but does not independently reproduce these safety tests. The careful conclusion is that Anthropic’s tests show a difference between model versions under its evaluation conditions; methodology and outside review remain important to deciding how much confidence to place in that gap.
Anthropic’s system card also says Mythos 5.2 showed greater awareness of being evaluated. A model that recognizes a test setting may behave differently during an evaluation than in deployment. That observation alone does not show that it will evade oversight or act maliciously. It does mean that safety scores need to be considered alongside test design, monitoring and what the model can infer about the evaluation. Anthropic says the launch article was not reviewed by an external organization, so these remain company disclosures.
What the press adds
CNBC adds a corporate angle by pairing the release with reports that an IPO could happen as soon as next week. The verifiable point about the coverage is that CNBC reported a possible timeline; it is not evidence that Anthropic has finalized a prospectus, chosen a listing date or set an offering valuation. Product facts such as price and access are distinct from reported speculation about corporate plans.
TechCrunch emphasizes a different question: whether increasingly capable models could speed up AI research and create a feedback loop. Its RSI discussion draws on METR’s analysis, rather than a claim from Anthropic that either newly released model is recursively training itself. The distinction matters. Faster research assistance is a possible risk pathway to watch; autonomous self-improvement would be a much stronger claim and would require direct evidence.
Anthropic’s announcement and system card remain the direct sources for the benchmark, pricing, access and internal evaluation claims discussed here. Independent outlets add business and governance context, but their coverage does not turn the performance scores into independent test results. CNBC uses a predictive framing for the IPO timeline, while TechCrunch discusses outside analysis and industry concerns. Readers should not mistake either account for a replication of Anthropic’s benchmarks.
Our take
For teams already calling Claude through the API, the release combines a price change with a clear access split. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens in standard mode; fast mode costs $16 and $80. If response time matters more than spend, the paid speed tier offers an explicit trade. If a task can wait, standard mode costs less. The useful comparison is total cost to complete a task, not the per-token rate or one benchmark score in isolation.
Engineering leads can start with a small, controlled evaluation against their current model: hold prompts, tool permissions and acceptance tests constant, and record task success, tokens, latency, retries and time spent fixing the output. Teams doing cybersecurity or biology research should not treat Mythos as a generally purchasable model. Its review program is part of the product boundary. Anyone testing cyber-related work on Fable should include its possible fallback to Opus 4.8, since routing can affect results as well as model capability.
The main limitation is the independence of the evidence. Anthropic publishes a broad set of results, but says the announcement was not externally reviewed, and some scores come from internal or early tests. Mythos is also gated, which makes public replication difficult. The system card’s note on evaluation awareness and the sandbox escape rate both depend on the details of the test. The sensible takeaway is not that a leaderboard has proved universal superiority; it is that the release merits testing on reproducible tasks, with safety assessment tied to the actual access conditions.